fix(diagnostics): make startup crashes self-reporting; cross-env for dev script

[julia-kafarska]

Follow-up to #39. That PR fixed the crash; this one makes the next one diagnosable.

Why

Three users independently root-caused #36/#37/#38 from OS crash reports, because the app recorded nothing itself. onnxruntime aborted natively inside InferenceSession.create — no JS error, no catch block, no chance to flush a buffered stream — and crashReporter was never started. %APPDATA%\Out Loud didn't exist afterwards, and the Windows event log was empty. From the user's side the app simply vanished ~3s after launch.

What

New electron/crash-log.ts: an append-only breadcrumb log shared by the main process and the TTS worker thread.

Every write is synchronous. That's the whole design constraint — the line has to be on disk before the next statement runs, so it survives an abort that never unwinds. Verified against process.abort() (SIGABRT, no flush): the breadcrumb persists.

console.log/warn/error are teed into it, so the breadcrumbs the app already emits (ONNX session providers: cpu, Model preloaded successfully) get captured in packaged builds, where stdout goes nowhere a user can reach.

The worker writes to the file directly rather than through its piped stdout. A native abort kills the thread with whatever is still sitting in the pipe — and the lost line is precisely the one identifying where it died. The main process passes the path via workerData; driving the worker standalone (as the website's sample generator does) just skips logging.

Also:

• crashReporter.start({ uploadToServer: false }) at module load, so native minidumps land in crashDumps. Local only — nothing is uploaded anywhere. Confirmed the Crashpad database is created on launch.
• The ONNX session parameters are now logged before the call that can abort, so even a truncated log names the providers, optimisation level and arch.
• child-process-gone, render-process-gone, uncaughtException, unhandledRejection handlers.
• Rotates at 2 MB keeping one previous file. All logging is best-effort and wrapped — diagnostics can never be what breaks startup.

Had this existed, #36 would have been a one-line paste:

=== session 2026-08-04T08:04:47.197Z version=2.1.1 electron=41.2.2 platform=darwin arch=arm64 packaged=false
...
[worker] Loaded embedded model: .../model_q8f16.onnx
[worker] [TTS Worker] Creating ONNX session: providers=[cpu] graphOptimizationLevel=all arch=arm64   <-- log ends here

Log location is printed at startup and is app.getPath("logs")/out-loud.log — %APPDATA%\Out Loud\logs\ on Windows, ~/Library/Logs/Out Loud/ on macOS.

Unrelated fix

electron:dev used Unix-only NODE_ENV=development electron ., which fails on Windows — also reported in #36. Routed through cross-env (added as a dev dependency).

Verification

• npm run check (lint, format, knip, typecheck) passes.
• Ran the app: single session header, main and worker breadcrumbs interleaved with timestamps, Crashpad database created.
• process.abort() survival test as described above.

Note: knip.config.js needed electron/crash-log.ts added to the entry list — the same treatment the other electron modules already get, since imports resolve to the committed .js and the .ts source would otherwise read as unused. The config already documents this.

🤖 Generated with Claude Code